[LPT] Added method to insert dequantization and avoid dq propagation to the shapeOf subgraps (#8214)

inference-engine/src/low_precision_transformations/include/low_precision/network_helper.hpp

@@ -210,6 +210,11 @@ public:
 
     static bool isPrecisionPreserved(const std::shared_ptr<ngraph::Node>& node);
 
+    static void insertDequantizationAfter(
+        const std::shared_ptr<Node>& originalNode,
+        const std::shared_ptr<Node>& dequantization,
+        const std::shared_ptr<Node>& newNode);
+
     static void replaceAttributeInNodes(
         std::shared_ptr<ngraph::Function> f,
         const std::string& name,

inference-engine/src/low_precision_transformations/include/low_precision/reduce_sum.hpp

@@ -17,7 +17,7 @@ namespace low_precision {
 class LP_TRANSFORMATIONS_API ReduceSumTransformation : public ReduceBaseTransformation {
 public:
     NGRAPH_RTTI_DECLARATION;
-    ReduceSumTransformation(const Params& params);
+    ReduceSumTransformation(const Params& params = Params());
     bool isPrecisionPreserved(std::shared_ptr<Node> reduce) const noexcept override;
     bool canBeTransformed(const TransformationContext& context, std::shared_ptr<Node> reduce) const override;
 

inference-engine/src/low_precision_transformations/src/add.cpp

@@ -227,7 +227,7 @@ bool AddTransformation::transform(TransformationContext& context, ngraph::patter
             ngraph::op::TemporaryReplaceOutputType(newAddOrSubtract, element::f32).get(),
             ngraph::op::TemporaryReplaceOutputType(multiplyEmptyPathValues, element::f32).get());
 
-        replace_node(add, newMultiply);
+        NetworkHelper::insertDequantizationAfter(add, newMultiply, newAddOrSubtract);
         NetworkHelper::copyInfo(add, newAddOrSubtract);
         ngraph::copy_runtime_info({ add, newMultiply }, newMultiply);
     }

inference-engine/src/low_precision_transformations/src/clamp.cpp

@@ -70,8 +70,7 @@ bool ClampTransformation::transform(TransformationContext& context, ngraph::patt
         replacement = std::make_shared<ngraph::opset1::Clamp>(newClamp->input_value(0), min, max);
     }
 
-    replace_node(newClamp, replacement);
+    replace_node_update_name(newClamp, replacement);
-    replacement->set_friendly_name(newClamp->get_friendly_name());
 
     element::Type outputClampType = dequantization.multiply ?
         dequantization.multiply->get_output_element_type(0) :

inference-engine/src/low_precision_transformations/src/concat.cpp

@@ -160,7 +160,7 @@ bool ConcatTransformation::transform(TransformationContext& context, ngraph::pat
         lastDequantization = multiply;
     }
 
-    replace_node(concat, lastDequantization);
+    NetworkHelper::insertDequantizationAfter(concat, lastDequantization, newConcat);
     NetworkHelper::copyInfo(concat, newConcat);
     updateOutput(context, lastDequantization, newConcat);
     return true;

inference-engine/src/low_precision_transformations/src/convolution.cpp

@@ -81,6 +81,7 @@ bool ConvolutionTransformation::transform(TransformationContext &context, ngraph
     convolution = NetworkHelper::separateInStandaloneBranch(convolution);
     FakeQuantizeDequantization dequantization = NetworkHelper::getDequantization(convolution);
 
+    std::shared_ptr<Node> newMultiplyAfter;
     {
         std::shared_ptr<opset1::Subtract> subtract;
         if (dequantization.subtract != nullptr) {
@@ -172,13 +173,13 @@ bool ConvolutionTransformation::transform(TransformationContext &context, ngraph
         }
         NetworkHelper::copyInfo(convolution, relaxedNewConvolution);
 
-        std::shared_ptr<ngraph::opset1::Multiply> newMultiplyAfter = std::make_shared<op::TypeRelaxed<opset1::Multiply>>(
+        newMultiplyAfter = std::make_shared<op::TypeRelaxed<opset1::Multiply>>(
             std::vector<element::Type>{ deqPrecision, deqPrecision },
             std::vector<element::Type>{ dequantization.multiply->get_output_element_type(0) },
             ngraph::op::TemporaryReplaceOutputType(relaxedNewConvolution, deqPrecision).get(),
             ngraph::op::TemporaryReplaceOutputType(newMultiplyAfterConst, deqPrecision).get());
 
-        replace_node(convolution, newMultiplyAfter);
+        NetworkHelper::insertDequantizationAfter(convolution, newMultiplyAfter, relaxedNewConvolution);
         convolution = newMultiplyAfter->input_value(0).get_node_shared_ptr();
 
         if (ov::is_type<opset1::Convert>(convolution->get_input_node_ptr(0))) {
@@ -242,7 +243,7 @@ bool ConvolutionTransformation::transform(TransformationContext &context, ngraph
             });
             NetworkHelper::copyInfo(convolution, newConvolution);
 
-            auto newMultiplyAfter = std::make_shared<opset1::Multiply>(
+            newMultiplyAfter = std::make_shared<opset1::Multiply>(
                 newConvolution,
                 foldConvert(
                     fold_reshape<opset1::Reshape>(
@@ -250,7 +251,7 @@ bool ConvolutionTransformation::transform(TransformationContext &context, ngraph
                         std::make_shared<opset1::Constant>(element::u64, Shape{ newScaleShape.size() }, newScaleShape),
                         false),
                     convolution->get_output_element_type(0)));
-            replace_node(convolution, newMultiplyAfter);
+            NetworkHelper::insertDequantizationAfter(convolution, newMultiplyAfter, newConvolution);
             convolution = newMultiplyAfter->input_value(0).get_node_shared_ptr();
         }
 
@@ -308,8 +309,7 @@ bool ConvolutionTransformation::transform(TransformationContext &context, ngraph
         }
     }
 
-    std::shared_ptr<ngraph::opset1::Multiply> finalDequantization = NetworkHelper::optimizeMultipliesAfter(
+    const auto finalDequantization = NetworkHelper::optimizeMultipliesAfter(newMultiplyAfter);
-        convolution->output(0).get_target_inputs().begin()->get_node()->shared_from_this());
     ngraph::copy_runtime_info({ convolution, finalDequantization }, finalDequantization);
     updateOutput(context, finalDequantization, convolution);
 

inference-engine/src/low_precision_transformations/src/convolution_backprop_data.cpp

@@ -98,6 +98,7 @@ bool ConvolutionBackpropDataTransformation::transform(TransformationContext &con
 
     convolutionBackpropData = NetworkHelper::separateInStandaloneBranch(convolutionBackpropData);
     FakeQuantizeDequantization dequantization = NetworkHelper::getDequantization(convolutionBackpropData);
+    std::shared_ptr<Node> newMultiplyAfter;
     {
         if (dequantization.subtract != nullptr) {
             NetworkHelper::optimizeSubtract(dequantization.subtract);
@@ -116,13 +117,13 @@ bool ConvolutionBackpropDataTransformation::transform(TransformationContext &con
             std::vector<element::Type>{deqPrecision, deqPrecision},
             std::vector<element::Type>{deqPrecision});
 
-        const auto newMultiplyAfter = std::make_shared<op::TypeRelaxed<opset1::Multiply>>(
+        newMultiplyAfter = std::make_shared<op::TypeRelaxed<opset1::Multiply>>(
             std::vector<element::Type>{ deqPrecision, deqPrecision },
             std::vector<element::Type>{ dequantization.multiply->get_output_element_type(0) },
             ngraph::op::TemporaryReplaceOutputType(relaxedConvolutionBackpropData, deqPrecision).get(),
             ngraph::op::TemporaryReplaceOutputType(newMultiplyAfterConst, deqPrecision).get());
+        NetworkHelper::insertDequantizationAfter(convolutionBackpropData, newMultiplyAfter, relaxedConvolutionBackpropData);
 
-        replace_node(convolutionBackpropData, newMultiplyAfter);
         convolutionBackpropData = newMultiplyAfter->get_input_node_shared_ptr(0);
         inputs[0] = convolutionBackpropData->get_input_node_ptr(0)->input_value(0);
         if (ov::is_type<opset1::Convert>(convolutionBackpropData->get_input_node_ptr(0))) {
@@ -154,15 +155,17 @@ bool ConvolutionBackpropDataTransformation::transform(TransformationContext &con
 
             auto inputs = convolutionBackpropData->input_values();
             inputs[1] = multiplyFromWeights->input_value(0);
-            auto newMultiplyAfter = std::make_shared<opset1::Multiply>(
+
-                convolutionBackpropData->clone_with_new_inputs(inputs),
+            const auto newconvolutionBackpropData = convolutionBackpropData->copy_with_new_inputs(inputs);
+            newMultiplyAfter = std::make_shared<opset1::Multiply>(
+                newconvolutionBackpropData,
                 foldConvert(
                     fold_reshape<opset1::Reshape>(
                         multiplyFromWeights->input_value(1),
                         std::make_shared<opset1::Constant>(element::u64, Shape{ newScaleShape.size() }, newScaleShape),
                         false),
                     convolutionBackpropData->get_output_element_type(0)));
-            replace_node(convolutionBackpropData, newMultiplyAfter);
+            NetworkHelper::insertDequantizationAfter(convolutionBackpropData, newMultiplyAfter, newconvolutionBackpropData);
             convolutionBackpropData = newMultiplyAfter->get_input_node_shared_ptr(0);
         }
 
@@ -202,8 +205,8 @@ bool ConvolutionBackpropDataTransformation::transform(TransformationContext &con
             convolutionBackpropData = newConvolution;
         }
     }
-    std::shared_ptr<ngraph::opset1::Multiply> finalDequantization = NetworkHelper::optimizeMultipliesAfter(
+
-            convolutionBackpropData->output(0).get_target_inputs().begin()->get_node()->shared_from_this());
+    const auto finalDequantization = NetworkHelper::optimizeMultipliesAfter(newMultiplyAfter);
     ngraph::copy_runtime_info({ convolutionBackpropData, finalDequantization }, finalDequantization);
     updateOutput(context, finalDequantization, convolutionBackpropData);
 

inference-engine/src/low_precision_transformations/src/fake_quantize_decomposition.cpp

@@ -230,7 +230,7 @@ std::tuple<std::shared_ptr<Node>, std::shared_ptr<Node>> decomposeFakeQuantize(
             deqPrecision,
             newFakeQuantizeLayer);
 
-        replace_node(layer, dequantization.multiply);
+        NetworkHelper::insertDequantizationAfter(layer, dequantization.multiply, newFakeQuantizeLayer);
 
         std::vector<std::shared_ptr<ngraph::Node>> sourceNodes{ layer };
         std::vector<std::shared_ptr<ngraph::Node>> targetNodes{ newFakeQuantizeLayer,  dequantization.multiply };

inference-engine/src/low_precision_transformations/src/mat_mul.cpp

@@ -167,7 +167,7 @@ bool MatMulTransformation::transform(TransformationContext &context, ngraph::pat
 
     newMultiply->set_friendly_name(newMatMul->get_friendly_name() + "/DequantizationMultiply");
 
-    replace_node(matMul, newMultiply);
+    NetworkHelper::insertDequantizationAfter(matMul, newMultiply, newMatMul);
     copy_runtime_info({ newMultiply, matMul }, newMultiply);
 
     updateOutput(context, newMultiply, newMatMul);

inference-engine/src/low_precision_transformations/src/mvn.cpp

@@ -159,7 +159,7 @@ bool MVNTransformation::transform(TransformationContext &context, ngraph::patter
         mvn->get_output_element_type(0));
     ngraph::copy_runtime_info({ mvn, newMultiply }, newMultiply);
 
-    replace_node(mvn, newMultiply);
+    NetworkHelper::insertDequantizationAfter(mvn, newMultiply, newMVN);
 
     updateOutput(context, newMultiply, newMVN);
     return true;

inference-engine/src/low_precision_transformations/src/network_helper.cpp

@@ -1108,7 +1108,7 @@ std::tuple<std::shared_ptr<Node>, std::shared_ptr<Node>> NetworkHelper::decompos
     dequantize->set_friendly_name(newFQ->get_friendly_name() + "/DequantizationMultiply");
     ngraph::copy_runtime_info({ newFQ, dequantize }, dequantize);
 
-    replace_node(fq, dequantize);
+    insertDequantizationAfter(fq, dequantize, newFQ);
 
     return std::make_tuple(newFQ, dequantize);
 }
@@ -1574,36 +1574,32 @@ NetworkHelper::InsertDequantizationResult NetworkHelper::moveDequantizationAfter
 
     assert(operation->get_output_size() == 1);
 
+    // we must have dequantization multiply
+    assert(dequantization.multiply != nullptr);
+
     OutputVector inputs = operation->input_values();
     const size_t dequantizationIndex = getChildInputIndex(dequantization.multiply, operation);
-    inputs[dequantizationIndex] = moveSubtract ?
+    inputs[dequantizationIndex] = (!moveSubtract && dequantization.subtract != nullptr) ?
-        dequantization.data :
+        dequantization.subtract :
-        (dequantization.subtract == nullptr ? dequantization.data : dequantization.subtract);
+        dequantization.data;
 
-    const std::shared_ptr<ngraph::Node> newOperation = operation->clone_with_new_inputs(inputs);
+    const auto newOperation = operation->clone_with_new_inputs(inputs);
     newOperation->set_friendly_name(operation->get_friendly_name());
     ngraph::copy_runtime_info(operation, newOperation);
 
-    auto op = std::dynamic_pointer_cast<ngraph::op::TypeRelaxedBase>(newOperation);
+    if (const auto op = std::dynamic_pointer_cast<ngraph::op::TypeRelaxedBase>(newOperation)) {
-    if (op != nullptr) {
+        op->set_overridden_output_type(updatePrecision ?
-        if (updatePrecision) {
+            newOperation->get_input_element_type(0) :
-            op->set_overridden_output_type(newOperation->get_input_element_type(0));
+            dequantization.multiplyConstant->get_element_type());
-        } else if (dequantization.multiply) {
+        newOperation->validate_and_infer_types();
-            op->set_overridden_output_type(dequantization.multiplyConstant->get_element_type());
-        } else if (dequantization.subtract) {
-            op->set_overridden_output_type(dequantization.subtractConstant->get_element_type());
-        }
-        std::dynamic_pointer_cast<ngraph::Node>(newOperation)->validate_and_infer_types();
     }
 
+    std::shared_ptr<Node> parent = newOperation;
+
     const element::Type deqPrecision = dequantization.multiplyConstant->get_element_type();
     const bool shouldConvert = (newOperation->get_output_element_type(0) != deqPrecision);
-
-    auto parent = newOperation;
     if (shouldConvert) {
-        const auto convertOutputPrecision = dequantization.convert != nullptr ?
+        const auto convertOutputPrecision = dequantization.convert ? dequantization.convert->get_element_type() : deqPrecision;
-            dequantization.convert->get_output_element_type(0) :
-            deqPrecision;
         parent = std::make_shared<opset1::Convert>(parent, convertOutputPrecision);
         ngraph::copy_runtime_info({ newOperation, parent }, parent);
     }
@@ -1619,12 +1615,9 @@ NetworkHelper::InsertDequantizationResult NetworkHelper::moveDequantizationAfter
             }
 
             parent = std::make_shared<op::TypeRelaxed<opset1::Subtract>>(
-                std::vector<element::Type>{element::f32, element::f32}, std::vector<element::Type>{ element::f32 },
+                element::TypeVector{ element::f32, element::f32 }, element::TypeVector{ element::f32 },
-                ngraph::op::TemporaryReplaceOutputType(parent, element::f32).get(),
+                op::TemporaryReplaceOutputType(parent, element::f32).get(),
-                ngraph::op::TemporaryReplaceOutputType(
+                op::TemporaryReplaceOutputType(foldConvert(dequantization.subtractConstant, parentPrecision), element::f32).get());
-                    dequantization.subtractConstant->output(0).get_element_type() == parentPrecision ?
-                        dequantization.subtractConstant :
-                        foldConvert(dequantization.subtractConstant->output(0), parentPrecision), element::f32).get());
             ngraph::copy_runtime_info({ newOperation, parent }, parent);
         } else {
             parent = std::make_shared<opset1::Subtract>(parent, dequantization.subtractConvert);
@@ -1633,23 +1626,21 @@ NetworkHelper::InsertDequantizationResult NetworkHelper::moveDequantizationAfter
     }
 
     if (dequantization.multiply != nullptr) {
-        auto multiplyConstant = dequantization.multiplyConstant;
         const element::Type parentPrecision = parent->get_output_element_type(0);
-        if (parentPrecision.bitwidth() < multiplyConstant->get_element_type().bitwidth()) {
+        if (parentPrecision.bitwidth() < dequantization.multiplyConstant->get_element_type().bitwidth()) {
             THROW_IE_LPT_EXCEPTION(*parent) <<
                 "unexpected precisions: on data " << parent->get_friendly_name() << ":" << parentPrecision <<
-                ", multiply dequantization constant " << multiplyConstant->get_friendly_name() << ":" << multiplyConstant->get_element_type();
+                ", multiply dequantization constant " << dequantization.multiplyConstant->get_friendly_name() << ":" <<
+                dequantization.multiplyConstant->get_element_type();
         }
 
         parent = std::make_shared<op::TypeRelaxed<opset1::Multiply>>(
-            opset1::Multiply(parent,
+            opset1::Multiply(parent, foldConvert(dequantization.multiplyConstant, parentPrecision)),
-                multiplyConstant->output(0).get_element_type() == parentPrecision ?
-                    multiplyConstant :
-                    foldConvert(multiplyConstant->output(0), parentPrecision)),
             dequantization.multiply->get_output_element_type(0));
         ngraph::copy_runtime_info({ newOperation, parent }, parent);
     }
-    replace_node(operation, parent);
+
+    insertDequantizationAfter(operation, parent, newOperation);
 
     if ((!moveSubtract) && (dequantization.convert != nullptr) && (dequantization.subtract != nullptr)) {
         // issue #43088
@@ -1878,6 +1869,21 @@ bool isDisabled(const std::shared_ptr<Node>& node) {
     }
     return false;
 }
+
+void NetworkHelper::insertDequantizationAfter(
+    const std::shared_ptr<Node>& originalNode,
+    const std::shared_ptr<Node>& dequantization,
+    const std::shared_ptr<Node>& newNode) {
+    replace_node(originalNode, dequantization);
+
+    // We do it to avoid dequantization propagation to the shapeOf subgraphs
+    for (const auto& input : dequantization->get_output_target_inputs(0)) {
+        if (const auto shapeOf = as_type_ptr<opset1::ShapeOf>(input.get_node()->shared_from_this())) {
+            const auto newShapeOf = shapeOf->clone_with_new_inputs({ newNode });
+            replace_node_update_name(shapeOf, newShapeOf);
+        }
+    }
+}
 } // namespace low_precision
 } // namespace pass
 } // namespace ngraph

inference-engine/src/low_precision_transformations/src/normalize_l2.cpp

@@ -140,7 +140,7 @@ bool NormalizeL2Transformation::transform(TransformationContext &context, ngraph
         ngraph::op::TemporaryReplaceOutputType(newNormalize, element::f32).get(),
         ngraph::op::TemporaryReplaceOutputType(newScalesConst, element::f32).get());
 
-    replace_node(normalize, newMultiply);
+    NetworkHelper::insertDequantizationAfter(normalize, newMultiply, newNormalize);
     ngraph::copy_runtime_info({ normalize, newMultiply }, newMultiply);
 
     updateOutput(context, newMultiply, newNormalize);

inference-engine/src/low_precision_transformations/src/split.cpp

@@ -109,6 +109,16 @@ bool SplitTransformation::transform(TransformationContext& context, ngraph::patt
         }
     }
 
+    // We do it to avoid dequantization propagation to the shapeOf subgraphs
+    for (size_t i = 0; i < replacement.size(); ++i) {
+        for (const auto& input : replacement[i].get_target_inputs()) {
+            if (const auto shapeOf = as_type_ptr<opset1::ShapeOf>(input.get_node()->shared_from_this())) {
+                const auto newShapeOf = shapeOf->clone_with_new_inputs({ newSplit->output(i) });
+                replace_node_update_name(shapeOf, newShapeOf);
+            }
+        }
+    }
+
     updateOutputs(context, lastNodes, newSplit);
     return true;
 }

inference-engine/tests/functional/inference_engine/lp_transformations/lpt_avoid_shapeof_propagation_test.cpp

@@ -0,0 +1,734 @@
+// Copyright (C) 2021 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include <string>
+#include <sstream>
+#include <memory>
+
+#include "low_precision/add.hpp"
+#include "low_precision/avg_pool.hpp"
+#include "low_precision/clamp.hpp"
+#include "low_precision/concat.hpp"
+#include "low_precision/convolution.hpp"
+#include "low_precision/convolution_backprop_data.hpp"
+#include "low_precision/depth_to_space.hpp"
+#include "low_precision/fake_quantize_decomposition.hpp"
+#include "low_precision/group_convolution.hpp"
+#include "low_precision/interpolate.hpp"
+#include "low_precision/mat_mul.hpp"
+#include "low_precision/max_pool.hpp"
+#include "low_precision/multiply.hpp"
+#include "low_precision/mvn.hpp"
+#include "low_precision/normalize_l2.hpp"
+#include "low_precision/pad.hpp"
+#include "low_precision/prelu.hpp"
+#include "low_precision/reduce_max.hpp"
+#include "low_precision/reduce_mean.hpp"
+#include "low_precision/reduce_min.hpp"
+#include "low_precision/reduce_sum.hpp"
+#include "low_precision/reshape.hpp"
+#include "low_precision/relu.hpp"
+#include "low_precision/squeeze.hpp"
+#include "low_precision/split.hpp"
+#include "low_precision/shuffle_channels.hpp"
+#include "low_precision/strided_slice.hpp"
+#include "low_precision/transpose.hpp"
+#include "low_precision/unsqueeze.hpp"
+#include "low_precision/variadic_split.hpp"
+
+#include "low_precision/network_helper.hpp"
+#include "lpt_ngraph_functions/common/builders.hpp"
+
+#include <gtest/gtest.h>
+#include <transformations/utils/utils.hpp>
+
+#include "common_test_utils/ngraph_test_utils.hpp"
+
+using namespace testing;
+using namespace ngraph;
+using namespace ngraph::pass;
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationAddTransformation) {
+    auto input1 = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto input2 = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+
+    auto convert1 = std::make_shared<opset1::Convert>(input1, element::f32);
+    auto convert2 = std::make_shared<opset1::Convert>(input2, element::f32);
+
+    auto mul1 = std::make_shared<opset1::Multiply>(convert1, opset1::Constant::create(element::f32, {}, { 2.f }));
+    auto mul2 = std::make_shared<opset1::Multiply>(convert2, opset1::Constant::create(element::f32, {}, { 4.f }));
+
+    auto add = std::make_shared<opset1::Add>(mul1, mul2);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(add);
+
+    auto result1 = std::make_shared<opset1::Result>(add);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input1, input2 });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::AddTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationAvgPoolTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto avgPool = std::make_shared<opset1::AvgPool>(mul, Strides{ 1, 1 }, Shape{ 1, 1 }, Shape{ 0, 0 }, Shape{ 2, 2 }, true);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(avgPool);
+
+    auto result1 = std::make_shared<opset1::Result>(avgPool);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::AvgPoolTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationClampTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto clamp = std::make_shared<opset1::Clamp>(mul, 0.0, 6.0);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(clamp);
+
+    auto result1 = std::make_shared<opset1::Result>(clamp);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ClampTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationConcatTransformation) {
+    auto input1 = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto input2 = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+
+    auto convert1 = std::make_shared<opset1::Convert>(input1, element::f32);
+    auto convert2 = std::make_shared<opset1::Convert>(input2, element::f32);
+
+    auto mul1 = std::make_shared<opset1::Multiply>(convert1, opset1::Constant::create(element::f32, {}, { 2.f }));
+    auto mul2 = std::make_shared<opset1::Multiply>(convert2, opset1::Constant::create(element::f32, {}, { 4.f }));
+
+    auto concat = std::make_shared<opset1::Concat>(OutputVector{ mul1, mul2 }, 1);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(concat);
+
+    auto result1 = std::make_shared<opset1::Result>(concat);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input1, input2 });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ConcatTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationConvolutionTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto weights = opset1::Constant::create(element::i8, { 6, 3, 1, 1 }, { 3 });
+    auto convertOnWeights = std::make_shared<opset1::Convert>(weights, element::f32);
+    auto mulOnWeights = std::make_shared<opset1::Multiply>(convertOnWeights, opset1::Constant::create(element::f32, {}, { 4.f }));
+
+    auto convolution = std::make_shared<opset1::Convolution>(
+        mul,
+        mulOnWeights,
+        ngraph::Strides{ 1, 1 },
+        ngraph::CoordinateDiff{ 0, 0 },
+        ngraph::CoordinateDiff{ 0, 0 },
+        ngraph::Strides{ 1, 1 });
+
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(convolution);
+
+    auto result1 = std::make_shared<opset1::Result>(convolution);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ConvolutionTransformation>();
+    m.run_passes(f);
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationConvolutionBackpropDataTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 8, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto weights = opset1::Constant::create(element::i8, { 8, 2, 1, 1 }, { 3 });
+    auto convertOnWeights = std::make_shared<opset1::Convert>(weights, element::f32);
+    auto mulOnWeights = std::make_shared<opset1::Multiply>(convertOnWeights, opset1::Constant::create(element::f32, {}, { 4.f }));
+
+    auto convolutionBackpropData = std::make_shared<opset1::ConvolutionBackpropData>(
+        mul,
+        mulOnWeights,
+        ngraph::Strides{ 1, 1 },
+        ngraph::CoordinateDiff{ 0, 0 },
+        ngraph::CoordinateDiff{ 0, 0 },
+        ngraph::Strides{ 1, 1 });
+
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(convolutionBackpropData);
+
+    auto result1 = std::make_shared<opset1::Result>(convolutionBackpropData);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ConvolutionBackpropDataTransformation>();
+    m.run_passes(f);
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationDepthToSpaceTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto d2s = std::make_shared<opset1::DepthToSpace>(mul, op::v0::DepthToSpace::DepthToSpaceMode::BLOCKS_FIRST);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(d2s);
+
+    auto result1 = std::make_shared<opset1::Result>(d2s);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::DepthToSpaceTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationFakeQuantizeDecompositionTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::f32, PartialShape{ 1, 3, 16, 16 });
+
+    ngraph::builder::subgraph::FakeQuantizeOnData fqValues{ 256ul, {}, {0.f}, {2.55f}, {0.f}, {2.55f} };
+    auto fakeQuantize = ngraph::builder::subgraph::makeFakeQuantize(input, element::f32, fqValues);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(fakeQuantize);
+
+    auto& outInfo = fakeQuantize->output(0).get_rt_info();
+    auto attribute = ngraph::pass::low_precision::make_shared_attribute<PrecisionsAttribute>(element::TypeVector{ element::u8, element::i8 });
+    auto attributeWrapper = std::make_shared<ngraph::VariantWrapper<std::shared_ptr<PrecisionsAttribute>>>(attribute);
+    outInfo.emplace(ngraph::VariantWrapper<std::shared_ptr<PrecisionsAttribute>>::type_info.name, attributeWrapper);
+
+    auto result1 = std::make_shared<opset1::Result>(fakeQuantize);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::FakeQuantizeDecompositionTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationGroupConvolutionTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 2 * 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto weights = opset1::Constant::create(element::i8, { 6, 3, 7, 7 }, { 2 });
+    auto convertOnWeights = std::make_shared<opset1::Convert>(weights, element::f32);
+    auto mulOnWeights = std::make_shared<opset1::Multiply>(convertOnWeights, opset1::Constant::create(element::f32, {}, { 4.f }));
+    auto reshapeConst = opset1::Constant::create(element::i32, { 5 }, { 2, 3, 3, 7, 7 });
+    auto reshapeOnWeights = std::make_shared<opset1::Reshape>(mulOnWeights, reshapeConst, true);
+
+    auto groupConvolution = std::make_shared<opset1::GroupConvolution>(
+        mul,
+        reshapeOnWeights,
+        ngraph::Strides{ 1, 1 },
+        ngraph::CoordinateDiff{ 0, 0 },
+        ngraph::CoordinateDiff{ 0, 0 },
+        ngraph::Strides{ 1, 1 });
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(groupConvolution);
+
+    auto result1 = std::make_shared<opset1::Result>(groupConvolution);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::GroupConvolutionTransformation>();
+    m.run_passes(f);
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationInterpolateTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto outShape = opset1::Constant::create(element::i32, { 4 }, { 1, 3, 18, 18});
+    op::v0::InterpolateAttrs attributes;
+    attributes.align_corners = false;
+    attributes.antialias = false;
+    attributes.axes = AxisSet{ 2, 3 };
+    attributes.mode = "nearest";
+    attributes.pads_begin = { 0 };
+    attributes.pads_end = { 0 };
+    auto interpolate = std::make_shared<opset1::Interpolate>(mul, outShape, attributes);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(interpolate);
+
+    auto result1 = std::make_shared<opset1::Result>(interpolate);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::InterpolateTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationMatMulTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 1024 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto weights = opset1::Constant::create(element::i8, { 2048, 1024 }, { 3 });
+    auto convertOnWeights = std::make_shared<opset1::Convert>(weights, element::f32);
+    auto mulOnWeights = std::make_shared<opset1::Multiply>(convertOnWeights, opset1::Constant::create(element::f32, {}, { 4.f }));
+
+    auto matmul = std::make_shared<opset1::MatMul>(mul, mulOnWeights, false, true);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(matmul);
+
+    auto result1 = std::make_shared<opset1::Result>(matmul);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::MatMulTransformation>();
+    m.run_passes(f);
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationMaxPoolTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto maxPool = std::make_shared<opset1::MaxPool>(mul, Strides{ 1, 1 }, Shape{ 1, 1 }, Shape{ 0, 0 }, Shape{ 2, 2 });
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(maxPool);
+
+    auto result1 = std::make_shared<opset1::Result>(maxPool);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::MaxPoolTransformation>();
+    m.run_passes(f);
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationMultiplyTransformation) {
+    auto input1 = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto input2 = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+
+    auto convert1 = std::make_shared<opset1::Convert>(input1, element::f32);
+    auto convert2 = std::make_shared<opset1::Convert>(input2, element::f32);
+
+    auto mul1 = std::make_shared<opset1::Multiply>(convert1, opset1::Constant::create(element::f32, {}, { 2.f }));
+    auto mul2 = std::make_shared<opset1::Multiply>(convert2, opset1::Constant::create(element::f32, {}, { 4.f }));
+
+    auto mul = std::make_shared<opset1::Multiply>(mul1, mul2);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(mul);
+
+    auto result1 = std::make_shared<opset1::Result>(mul);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input1, input2 });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::MultiplyTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationMVNTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto MVN = std::make_shared<op::TypeRelaxed<op::v0::MVN>>(mul);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(MVN);
+
+    auto result1 = std::make_shared<opset1::Result>(MVN);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::MVNTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationNormalizeL2Transformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axes = opset1::Constant::create(element::i32, { 2 }, { 2, 3 });
+    auto normalize = std::make_shared<opset1::NormalizeL2>(mul, axes, 0.01, ov::op::EpsMode::ADD);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(normalize);
+
+    auto result1 = std::make_shared<opset1::Result>(normalize);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::NormalizeL2Transformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationPadTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto pads_begin = opset1::Constant::create(element::i32, { 4 }, { 0, 0, 1, 1 });
+    auto pads_end = opset1::Constant::create(element::i32, { 4 }, { 0, 0, 1, 1 });
+    auto pad = std::make_shared<opset1::Pad>(mul, pads_begin, pads_end, op::PadMode::CONSTANT);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(pad);
+
+    auto result1 = std::make_shared<opset1::Result>(pad);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::PadTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationPReluTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto slope = opset1::Constant::create(element::f32, { 1, 3, 1, 1 }, { 0.01f });
+    auto prelu = std::make_shared<opset1::PRelu>(mul, slope);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(prelu);
+
+    auto result1 = std::make_shared<opset1::Result>(prelu);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::PReluTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationReduceMaxTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axes = opset1::Constant::create(element::i32, { 2 }, { 2, 3 });
+    auto reduce = std::make_shared<opset1::ReduceMax>(mul, axes);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(reduce);
+
+    auto result1 = std::make_shared<opset1::Result>(reduce);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ReduceMaxTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationReduceMeanTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axes = opset1::Constant::create(element::i32, { 2 }, { 2, 3 });
+    auto reduce = std::make_shared<opset1::ReduceMean>(mul, axes);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(reduce);
+
+    auto result1 = std::make_shared<opset1::Result>(reduce);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ReduceMeanTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationReduceMinTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axes = opset1::Constant::create(element::i32, { 2 }, { 2, 3 });
+    auto reduce = std::make_shared<opset1::ReduceMin>(mul, axes);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(reduce);
+
+    auto result1 = std::make_shared<opset1::Result>(reduce);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ReduceMinTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationReduceSumTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axes = opset1::Constant::create(element::i32, { 2 }, { 2, 3 });
+    auto reduce = std::make_shared<opset1::ReduceSum>(mul, axes);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(reduce);
+
+    auto result1 = std::make_shared<opset1::Result>(reduce);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ReduceSumTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationReshapeTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto outShape = opset1::Constant::create(element::i32, { 3 }, { 1, 3, -1 });
+    auto reshape = std::make_shared<opset1::Reshape>(mul, outShape, true);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(reshape);
+
+    auto result1 = std::make_shared<opset1::Result>(reshape);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ReshapeTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationReluTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto relu = std::make_shared<opset1::Relu>(mul);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(relu);
+
+    auto result1 = std::make_shared<opset1::Result>(relu);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ReluTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationSqueezeTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axes = opset1::Constant::create(element::i32, { 1 }, { 0 });
+    auto squeeze = std::make_shared<opset1::Squeeze>(mul, axes);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(squeeze);
+
+    auto result1 = std::make_shared<opset1::Result>(squeeze);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::SqueezeTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationSplitTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axis = opset1::Constant::create(element::i32, {}, { 1 });
+    auto split = std::make_shared<opset1::Split>(mul, axis, 3);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(split);
+
+    auto result1 = std::make_shared<opset1::Result>(split->output(0));
+    auto result2 = std::make_shared<opset1::Result>(split->output(1));
+    auto result3 = std::make_shared<opset1::Result>(split->output(2));
+    auto result4 = std::make_shared<opset1::Result>(shapeOf->output(0));
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2, result3, result4 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::SplitTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationShuffleChannelsTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto shuffleChannels = std::make_shared<opset1::ShuffleChannels>(mul);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(shuffleChannels);
+
+    auto result1 = std::make_shared<opset1::Result>(shuffleChannels);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::ShuffleChannelsTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationStridedSliceTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto beginParam = ngraph::op::Constant::create(ngraph::element::i64, ngraph::Shape{ 4 }, { 0, 0, 0, 0 });
+    auto endParam = ngraph::op::Constant::create(ngraph::element::i64, ngraph::Shape{ 4 }, { 1, 2, 1, 1 });
+    auto stridesParam = ngraph::op::Constant::create(ngraph::element::i64, ngraph::Shape{ 4 }, { 1, 1, 1, 1 });
+    auto stridedSlice = std::make_shared<ngraph::opset1::StridedSlice>(
+        mul, beginParam, endParam, stridesParam,
+        std::vector<std::int64_t>{ 1, 0, 1, 1 },
+        std::vector<std::int64_t>{ 1, 0, 1, 1 });
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(stridedSlice);
+
+    auto result1 = std::make_shared<opset1::Result>(stridedSlice);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::StridedSliceTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationTransposeTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto constant = ngraph::op::Constant::create(ngraph::element::i64, ngraph::Shape{ 4 }, { 0, 1, 3, 2 });
+    auto transpose = std::make_shared<ngraph::opset1::Transpose>(mul, constant);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(transpose);
+
+    auto result1 = std::make_shared<opset1::Result>(transpose);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::TransposeTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationUnsqueezeTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axes = opset1::Constant::create(element::i32, { 1 }, { 3 });
+    auto unsqueeze = std::make_shared<opset1::Unsqueeze>(mul, axes);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(unsqueeze);
+
+    auto result1 = std::make_shared<opset1::Result>(unsqueeze);
+    auto result2 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::UnsqueezeTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}
+
+TEST(LPT, AvoidDequantizationToShapeOfPropagationVariadicSplitTransformation) {
+    auto input = std::make_shared<opset1::Parameter>(element::u8, PartialShape{ 1, 3, 16, 16 });
+    auto convert = std::make_shared<opset1::Convert>(input, element::f32);
+    auto mul = std::make_shared<opset1::Multiply>(convert, opset1::Constant::create(element::f32, {}, { 2.f }));
+
+    auto axis = opset1::Constant::create(element::i32, {}, { 1 });
+    auto lengths = opset1::Constant::create(element::i32, { 2 }, { 1, 2 });
+    auto variadicSplit = std::make_shared<opset1::VariadicSplit>(mul, axis, lengths);
+    auto shapeOf = std::make_shared<opset1::ShapeOf>(variadicSplit->output(0));
+
+    auto result1 = std::make_shared<opset1::Result>(variadicSplit->output(0));
+    auto result2 = std::make_shared<opset1::Result>(variadicSplit->output(1));
+    auto result3 = std::make_shared<opset1::Result>(shapeOf);
+
+    auto f = std::make_shared<Function>(ResultVector{ result1, result2, result3 }, ParameterVector{ input });
+    pass::Manager m;
+    m.register_pass<pass::low_precision::VariadicSplitTransformation>();
+    m.run_passes(f);
+
+    auto dqBeforeShapeOf = low_precision::NetworkHelper::getDequantization(result2->get_input_node_shared_ptr(0));
+    ASSERT_TRUE(dqBeforeShapeOf.empty());
+}